AI racks approach megawatt power levels as SiC, GaN, silicon divide power delivery roles

Megawatt-class AI racks represent a paradigm rupture, not an incremental upgrade. Silicon MOSFETs have hit a hard physical ceiling at 800V bus conversion, forcing a rigid tri-material split: SiC owns primary high-voltage conversion, GaN captures high-frequency secondary stages, and silicon retreats to protection logic and low-voltage current paths. This is not a substitutable menu—it is a structural division of labor. Upstream, the 6-to-8-inch SiC substrate transition accelerates while GaN-on-Si epitaxy becomes the new chokepoint. China's gallium and germanium export controls have structurally raised the marginal cost of GaAs-based GaN, pushing the industry toward GaN-on-Si and eroding the legacy substrate advantage held by manufacturers in Japan and Taiwan, China. Competitively, Infineon and onsemi are locking in SiC module share through customer co-design, while NVIDIA's 800V HVDC reference architecture is effectively dictating upstream specifications from the demand side. Pure-play GaN vendors face a consolidation window as IDM players absorb the value chain. Core call for the next 12-24 months: power semiconductors shift from invisible enabler to the number-one capacity constraint in AI infrastructure buildout. Hybrid SiC-plus-GaN-plus-silicon integration becomes the default datacenter PSU architecture by 2026, and single-material pure-plays face a valuation reset.